Deep Learning on Automatic ICD Coding for Clinical Decision Support Systems: Review

The International Code of Disease is a coding method for diagnosing diseases using a clinical text with many benefits, such as insurance claims, medical bills, and hospital finances. However, this coding method is a high-complexity process with a high risk of failure. Recently, researchers have explored the use of Natural Language Processing and various Classification Models to develop Automatic ICD Coding to encounter the manual coding process by assisting doctors and medical staff. This literature review aims to analyze and provide a comprehensive overview of Automatic ICD Coding research. We conducted the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) diagram as a guiding principle. We created keywords to search relevant papers between 2021 and 2024 on the SCOPUS database. Our review included 25 relevant papers to be reviewed. This literature review shows that the 10th version of the ICD code is the most popular version, with 14 studies, and the highest number of categories is 17080 categories. Our analysis concluded that MIMIC-III, a publicly available dataset, is the most frequently utilized dataset in this research field. Researchers have used English as a dataset language to develop automatic ICD coding systems. Researcher have conducted their novel classification model to be implemented on Automatic ICD Coding Systems. Future research is able to focus on applying deep learning methods to automatic ICD coding systems using datasets in different languages, ICD versions & categories. Addressing these research gaps is crucial for meeting the evolving needs of clinical and healthcare systems.